Non Asymptotic Performance of Some Markov Chains Order Estimators

TL;DR

This paper compares the non-asymptotic performance of various Markov chain order estimators, highlighting the superior efficiency of the GDL estimator over AIC, BIC, and EDC, especially for small samples.

Contribution

It introduces and analyzes the GDL estimator, demonstrating its strong consistency and improved performance over traditional estimators in finite sample scenarios.

Findings

GDL outperforms AIC, BIC, and EDC in small samples.

GDL is strongly consistent and more efficient.

Different structural properties of estimators affect their performance.

Abstract

In what follows we study non asymptotic behavior of different well known estimators AIC(\cite{Tong}), BIC(\cite{Schwarz}) and EDC(\cite{Zhao,Dorea}) in contrast with the Markov chain order estimator, named as Global Depency Level - GDL(\cite{Baigorri}). The estimator GDL, is based on a different principle which makes it behave in a quite different form. It is strongly consistent and more efficient than AIC(inconsistent), outperforming the well established and consistent BIC and EDC, mainly on relatively small samples. The estimators mentioned above mainly consist in the evaluation of the Markov chain's sample by different multivariate deterministic functions. The log likelihood approach or the GDL approach, shall be analysed exhibiting different structural properties. It will become clear the intimate differences existing between the variance of both estimators, which induce quite dissimilar performance, mainly for samples of moderated sizes.